Granular ball-based partial label feature selection via fuzzy correlation and redundancy

Abstract

Partial label learning is a weakly supervised framework in which each training sample is associated with a set of candidate labels, but only one among them is the true label. Feature selection is a technique for enhancing the ability of learning models to generalize effectively. However, a challenging problem in feature selection for partial label learning is the impact of ambiguous candidate labels. To address this, this paper proposes a granular ball-based partial label feature selection method via fuzzy correlation and redundancy. Firstly, the paper utilizes granular ball computing to obtain two granular ball sets that respectively reflect the supervision information from candidate and non-candidate labels. The relative density between two granular ball sets is used to obtain labeling confidence which can identify the ground-truth labels. Then, a novel fuzzy entropy is defined by combining consistency in the granular ball with fuzzy information entropy. Additionally, fuzzy mutual information is derived by considering the fuzzy entropy and the fuzzy similarity constrained by granular ball radius. Fuzzy correlation and redundancy is measured by granular ball-based fuzzy mutual information. A heuristic search strategy is used to rank the features according to the principle of maximizing relevance and minimizing redundancy. Finally, experimental results on five real-world datasets and eight controlled UCI datasets show that the proposed method obtains superior performance than other compared methods.